{% comment %}
Copyright (C) The Arvados Authors. All rights reserved.

SPDX-License-Identifier: CC-BY-SA-3.0
{% endcomment %}

h2. Runtime constraints

Runtime constraints restrict the container's access to compute resources and the outside world (in addition to its explicitly stated inputs and output).

table(table table-bordered table-condensed).
|_. Key|_. Type|_. Description|_. Notes|
|ram|integer|Number of ram bytes to be used to run this process.|Optional. However, a ContainerRequest that is in "Committed" state must provide this.|
|vcpus|integer|Number of cores to be used to run this process.|Optional. However, a ContainerRequest that is in "Committed" state must provide this.|
|keep_cache_disk|integer|When the container process accesses data from Keep via the filesystem, that data will be cached on disk, up to this amount in bytes.|Optional. If your cluster is configured to use a disk cache by default, the default size will match your @ram@ constraint, bounded between 2GiB and 32GiB.|
|keep_cache_ram|integer|When the container process accesses data from Keep via the filesystem, that data will be cached in memory, up to this amount in bytes.|Optional. If your cluster is configured to use a RAM cache by default, the administrator sets a default cache size.|
|API|boolean|When set, ARVADOS_API_HOST and ARVADOS_API_TOKEN will be set, and container will have networking enabled to access the Arvados API server.|Optional.|
|gpu|object|Request GPU support, see below|Optional.|
|cuda|object|Old way to request CUDA GPU support, included for backwards compatability only.  Use the 'gpu' field instead.|Deprecated.|

h3. GPU support

table(table table-bordered table-condensed).
|_. Key|_. Type|_. Description|_. Notes|
|stack|string|One of 'cuda' or 'rocm' to request Nvidia or AMD GPU support.||
|device_count|int|Number of GPUs to request.|Count greater than 0 enables GPU support.|
|driver_version|string|Minimum driver version, in "X.Y" format.|Required when device_count > 0|
|hardware_target|array of strings|For CUDA: a single item with minimum CUDA hardware capability, in "X.Y" format, or multiple items listing CUDA specific hardware capability versions, one of which must be an exact match on the compute node the container is scheduled on.
For ROCm: A list of one or more hardware targets (e.g. gfx1100) corresponding to the GPU architectures supported by the container.  To be scheduled, at least one item in this list must match the @HardwareTarget@ of one of the cluster's @InstanceTypes@.|Required when device_count > 0|
|vram|int|Amount of VRAM to request, in bytes.||

h3. CUDA support (deprecated)

Note.  This API is deprecated.  Use the 'gpu' API instead.

table(table table-bordered table-condensed).
|_. Key|_. Type|_. Description|_. Notes|
|device_count|int|Number of GPUs to request.|Count greater than 0 enables CUDA GPU support.|
|driver_version|string|Minimum CUDA driver version, in "X.Y" format.|Required when device_count > 0|
|hardware_capability|string|Minimum CUDA hardware capability, in "X.Y" format.|Required when device_count > 0|